compounding-log
Skill0 starsChanged 46 days agoMITView on GitHub
Make the whole venture system compound — capture what worked and failed across products and fold concrete improvements back into the recipe-book skills, so each week is faster and sharper. Use this skill for the weekly review/retro, when the user says "what did we learn", "update the recipe book", "compound the learnings", "calibrate the gates", "weekly compounding", or at the end of a build/launch cycle. This is the M2 meta-skill of the Summer Solstice recipe book — the business-level twin of compound engineering's `ce-compound` (which compounds one codebase). Without it, the sprint is 12 disconnected attempts; with it, it's a learning system.
In this repository9 other files it gives its agents.
- AGENTS.mdAGENTS.md
- build-handoffSkill
- demand-confidence-gateSkill
- grow-sell-killSkill
- launch-distributionSkill
- opportunity-scoringSkill
- opportunity-sourcingSkill
- positioning-teardownSkill
- seed-market-selectionSkill
The file
We split this skill into the rules it gives an agent
6 rules. About 0 are common, 5 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 19
Run weekly (the Sunday "Compound" step of the conveyor).
- Distinctive: line 52
update the channel bullseye defaults in Launch (S11),
- Distinctive: line 53
add a new skill if the same manual work recurred three times (the agent-native latent-demand rule).
- Very specific: line 62
Rubric-sync check (run every week, non-negotiable): verify rubric.json agrees with all mirrors — Opportunity-Scorecard.xlsx weights/gates, …
- Distinctive: line 69
Write a short "what changed in the recipe book this week and why" entry.
- Distinctive: line 87
Recency bias: Use the full history, not just last week.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Compounding Log (M2)
The engine behind "the recipe book gets better every week." ce-compound makes a single codebase compound; this makes the whole system compound — sourcing, scoring, demand-testing, building, launching — by comparing what we predicted to what actually happened and editing the skills accordingly.
Cadence
Run weekly (the Sunday "Compound" step of the conveyor). Can be wired to a scheduled task. Also run after any product hits a milestone or dies.
Step 1 — Log the cycle (per product)
Append to compounding-log.md (and keep sourcing-log.md in sync):
- Seed (S0): the seed + its market-attractiveness score.
- Sourcing (S1): which probes/sources yielded the candidate.
- Scoring (S2): the Scorecard scores + verdict.
- Demand gate (S5): the predicted confidence (High/Med/Low) — to be compared with reality.
- Build (S6): notable build notes + the CLI-leverage decision.
- Launch (S11): channel(s) + result.
- Traction (S15): current alive metric / MRR / retention.
Step 2 — Calibration loops (the heart)
Compare predicted vs actual and turn the gaps into edits:
- S5 confidence vs outcome: did "High" ideas actually get usage/revenue? If High-confidence ideas keep dying, the kill bar is too loose — tighten it. If killed ideas later succeed elsewhere, it's too tight.
- S0 attractiveness vs winners: which seed lenses produced products that got traction? Re-weight the lenses toward those.
- S2 scores vs traction: which Scorecard criteria actually predicted success? Re-weight; retire criteria that don't discriminate.
- S1 probes vs yield: which probe patterns surfaced real candidates? Promote them; drop dead ones.
- S11 channels vs activation: which channels converted for which product types? Update the channel bullseye defaults.
This is Pincus's point made systematic: sharpen instinct with feedback, don't trust it blindly.
Step 3 — Fold improvements back into the skills
For each calibration finding, produce a concrete edit and apply it:
- re-weight the rubric — edit
rubric.jsonfirst (via S2, its keeper), then regenerate the mirrors (xlsx, OS-doc §7, S2's tables), - promote/retire probe patterns in Opportunity Sourcing (S1),
- adjust the Demand Confidence Gate kill bar (S5),
- update the channel bullseye defaults in Launch (S11),
- add a new skill if the same manual work recurred three times (the agent-native latent-demand rule).
Human gate: propose the edits with rationale; the human approves significant changes (light touch — small re-weights can be auto-applied, structural changes get a yes).
Anti-overfit guardrails (External Review #01): rubric weight edits require n≥3 logged product outcomes pointing the same way; structural changes (gates or criteria added/removed) require n≥5 or an unambiguous external fact (e.g., a channel's payment rail doesn't exist). One rehearsal is a flag, not a trend — log it and wait.
Rubric-sync check (run every week, non-negotiable): verify rubric.json agrees with all mirrors —
Opportunity-Scorecard.xlsx weights/gates, operating-system doc §7, and the tables in the
opportunity-scoring skill. Also verify no other skill restates the rubric (S1 once drifted to v1
silently — that's how External Review #01 found it). Any mismatch is a bug: fix it now and log it.
Step 4 — Publish the recipe-book changelog
Write a short "what changed in the recipe book this week and why" entry. This is the visible proof the system is compounding — and the running record of how the playbook was earned.
Output format
COMPOUNDING LOG — week of <date>
Cycles logged: <products>
Calibration findings:
• <predicted vs actual> → <edit>
Edits applied / proposed: <skill → change → rationale> [✓ applied | 🔲 needs approval | ⏸ awaiting min-n]
Rubric-sync check: <rubric.json vs xlsx vs OS-doc vs S2 tables — PASS/FAIL + fixes>
Recipe-book changelog: <1–3 lines>
Next-week bias: <what we'll do differently>Anti-patterns
- Logging without acting. A log that doesn't change a skill is journaling, not compounding.
- Recency bias. Use the full history, not just last week.
- Over-fitting to one product. Re-weight on patterns across several, not a single outlier.
- Skipping the human gate on structural changes — small re-weights auto-apply; rubric/skill restructures get a yes.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

